Problem description
I am trying to implement a custom Keras model (StreamingModel) that includes a custom layer (CircularBufferLayer). The following error was encountered while compiling the model:
NotImplementedError: Exception encountered when calling StreamingModel.call().
Could not automatically infer the output shape / dtype of 'streaming_model' (of type StreamingModel).
Either the `StreamingModel.call()` method is incorrect, or you need to implement the `StreamingModel.compute_output_spec() / compute_output_shape()` method.
Error encountered: Model StreamingModel does not have a `call()` method implemented.
I have checked the code, but I am not sure how to correctly implement the call() method or whether it is necessary to implement the compute_output_spec() method. The following is the relevant code and error log:
import tensorflow as tf
class CircularBufferLayer(tf.keras.layers.Layer):
def __init__(self, num_features, buffer_size, stride, **kwargs):
super().__init__(**kwargs)
self.num_features = num_features
self.buffer_size = buffer_size
self.stride = stride
self.gradient_scale = 0.1
self.buffer = self.add_weight(name='buffer', shape=(1, buffer_size, num_features),
initializer='zeros', trainable=False, dtype=tf.float32)
self.call_count = self.add_weight(name='call_count', shape=(), initializer='zeros',
dtype=tf.int32, trainable=False)
self.total_call_count = self.add_weight(name='total_call_count', shape=(), initializer='zeros',
dtype=tf.int32, trainable=False)
def call(self, inputs):
scaled_input = tf.multiply(inputs, self.gradient_scale)
new_buffer = tf.concat([scaled_input, self.buffer[:, :-1]], axis=1)
self.buffer.assign(new_buffer)
return self.buffer
class StreamingModel(tf.keras.Model):
def call(self, inputs):
x, _ = super().call(inputs)
return x
buffer_layer = CircularBufferLayer(num_features=64, buffer_size=10, stride=1)
model = StreamingModel()
input_shape = (None, 10, 64) # (batch_size, sequence_length, num_features)
inputs = tf.keras.Input(shape=input_shape[1:])
outputs = model(inputs)
model = tf.keras.Model(inputs=inputs, outputs=outputs)
model.compile(
optimizer='adam',
loss='mse',
metrics=['mae']
)
x_train = tf.random.normal((100, 10, 64))
y_train = tf.random.normal((100, 10, 64))
history = model.fit(
x_train, y_train,
batch_size=32,
epochs=10,
validation_split=0.2
)
model.save('streaming_model.h5')
error log:
2025-03-19 16:09:45.545622: I tensorflow/core/util/port.cc:153] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
2025-03-19 16:09:46.283611: I tensorflow/core/util/port.cc:153] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
2025-03-19 16:09:48.149403: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 AVX_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
WARNING:tensorflow:From D:\project\DLComplier\.venv\Lib\site-packages\keras\src\backend\tensorflow\core.py:216: The name tf.placeholder is deprecated. Please use tf.compat.v1.placeholder instead.
Traceback (most recent call last):
File "D:\project\DLComplier\.venv\statiblity.py", line 37, in <module>
outputs = model(inputs)
^^^^^^^^^^^^^
File "D:\project\DLComplier\.venv\Lib\site-packages\keras\src\utils\traceback_utils.py", line 122, in error_handler
raise e.with_traceback(filtered_tb) from None
File "D:\project\DLComplier\.venv\statiblity.py", line 27, in call
x, _ = super().call(inputs) # 假设另一个分支被截断
^^^^^^^^^^^^^^^^^^^^
NotImplementedError: Exception encountered when calling StreamingModel.call().
Could not automatically infer the output shape / dtype of 'streaming_model' (of type StreamingModel). Either the `StreamingModel.call()` method is incorrect, or you need to implement the `StreamingModel.compute_output_spec() / compute_output_shape()` method. Error encountered:
Model StreamingModel does not have a `call()` method implemented.
Arguments received by StreamingModel.call():
• args=('<KerasTensor shape=(None, 10, 64), dtype=float32, sparse=False, name=keras_tensor>',)
• kwargs=<class 'inspect._empty'>
environmental information
TensorFlow version: 2.18.0
Python version: 3.12.6
Operating System: Windows 11
GPU/CPU: CPU
Supplementary explanation：Is this a new compliation error?